Deep Correlation Filter based Real-Time Tracker

Visual tracking is a challenging task in computer vision. Recently correlation filter based trackers have gained much attention due to their high efficiency and impressive performance. Several methods have been developed to utilize deep features extracted from Convolutional Neural Networks (CNNs) for correlation filter tracking. Despite their success, most of these approaches suffer from low tracking speed due to high computation burden of the deep models. In this paper, we propose a CNN based real-time deep tracker for robust visual tracking. We first exploit hierarchical deep features as target representation, which can better distinguish the target from the backgrounds in the presence of complex situations. Before applied to learn correlation filters, the channel number of hierarchical features is reduced by the principal component analysis (PCA). This method can decrease both feature redundancy and computation. Then we construct a reliable map with features from the last convolutional layer as they encode the most semantic information. The reliable map is used to constraint the searching area where the target most likely exist. To further handle long-term tracking, we improve the model discrimination capability by introducing the original template into the current correlation filter model. As the target movement is usually small between two adjacent frames, we reuse the deep features from location step to update current model. The features are reused by making a circular shift with position displacement of the target, which increases the tracking speed significantly. Extensive experimental results on large benchmarks show that our proposed tracker can perform at real-time and achieves the state-of-the-art performance.